AI-powered space robotics answers that arithmetic directly. Software carried on the vehicle converts long silences into productive work by reading terrain, choosing science targets, and rebuilding its own schedule. The pressure increases with distance, which is why deep-space exploration has become the strongest argument for onboard intelligence.1,2
Autonomy at Work on Mars
Perseverance drove 17.7 km during its first two years, sealed 23 sample tubes, and deposited 10 of them at a site chosen for later retrieval. By Sol 753 in April 2023, 88.7% of its driving had been planned or mapped autonomously, and it had broken the old single-day distance record more than 22 times.2
Three systems carry that workload. AutoNav builds terrain maps and selects safe paths while the wheels keep turning. AEGIS picks scientifically interesting rocks after a drive, images them, and fires a laser spectrometer at them. The OnBoard Planner reschedules activities to conserve energy and stretch each day further.2
Autonomy also protects the link home. AutoNav can turn the rover toward a communication heading, and when the current spot is unsafe, it drives about 4 more meters until a safe orientation appears. The rover updates its pose at 8 Hz and tracks the Sun to correct accumulated attitude error.2
Language Models Join the Route Planners
In December 2025, engineers at the Jet Propulsion Laboratory used vision-language models from Anthropic to plan two Perseverance drives along the rim of Jezero Crater. The rover covered 210 m on December 8 and 246 m on December 10, following routes produced without human input.1
The models worked from the same evidence people use, including high-resolution orbital imagery from the Mars Reconnaissance Orbiter, slope data drawn from digital elevation models, and the archive of past surface operations. They identified bedrock, outcrops, hazardous boulder fields, and sand ripples, then generated a continuous path with waypoints.1
Human planners normally sketch waypoints no more than 100 m apart, since each segment must be defensible before it is transmitted. Shifting that judgment into software compresses the planning cycle, and the same compression would let a lunar vehicle cover far more ground between conversations with its operators.1
Planning Across Two Scales
Rover navigation divides into global and local layers. Control centers on Earth generate long traverses from orbital topographic maps, optimizing for route length, energy use, or lighting conditions. Slope, surface roughness, communication constraints, and vehicle limits all enter the calculation before a single wheel turns.3
Orbital maps stay too coarse to resolve individual obstacles, and they say little about shifting illumination or temperature. Local planning fills that gap by reconstructing terrain from onboard stereo imagery and replanning around hazards as they appear. Curiosity, Opportunity, and China's Zhurong rover all depended on this reactive layer.3
Designers work inside a tight budget. A local planner has to respond quickly using modest processors and a small power supply, which rules out elaborate computation. Machine learning methods are moving into this layer, and the trade-off between accuracy and available onboard resources governs which ones survive review.3
Learning to Drive Faster
European research targets speed. Curiosity and Perseverance average about 4.2 cm/s and rarely exceed 100 m in a day, partly because the classic sense, model, plan, act cycle forces repeated stops for terrain analysis. Sample return campaigns that need long traverses feel this limit acutely.4
The FASTNAV system replaces that rhythm with multi-range perception and two operating modes, a cautious one named RAPID and an opportunistic one that relies on long-range hazard detection. A learned Far Obstacle Detector provides warnings at configurable distances, enabling sustained speeds above 1 m/s.4
Flight-relevant hardware makes such perception credible. A MobileNet detector runs in 4 ms at below two watts on a space-qualified Movidius processor, and semantic segmentation with DeepLabV3+ completes in ten milliseconds on a Xilinx FPGA that is available in radiation-hardened variants.4
Training data stays scarce, so teams manufacture it. Labeled corpora include 3,000 rock images from a European coastal test site, roughly 35,000 Mars images carrying about 326,000 semantic labels, and 2,784 rendered lunar scenes with per-pixel masks generated for synthetic training.4
Lunar Work Keeps People in the Loop
The VIPER project illustrates a more conservative balance. NASA planned to drive the rover remotely at the lunar south pole, with hazard detection software running on Earth rather than aboard the vehicle. The software marks dangerous ground in camera images and passes its conclusions to human operators.5
Light behaves awkwardly near the pole, where the Sun hangs low, and shadows stretch for hundreds of meters. Computer vision estimates local slope from pixel brightness, Sun position, camera pointing, and the reflective behavior of regolith, producing terrain models detailed enough to predict how those shadows move.5
A planner named SHERPA uses shadow maps to make decisions under uncertainty, marking the first time this technique has been used on a space mission. Sparse training data and the novelty of remote lunar driving explain the deliberate decision to keep humans holding final authority.5
Repair and Assembly in Orbit
Servicing spacecraft demands comparable judgment closer to home. Recent work on small satellite servicing recommends onboard AI for real-time perception, fault diagnosis, and adaptive task planning. This is organized under a system known as supervised autonomy, where ground teams intervene only at critical decision points.6
Download the PDF of this page here
Proximity operations allow little delay. Learning-based collision avoidance lets a servicing vehicle process sensor data, predict trajectories, and adjust its path without waiting for a command. This capability is increasingly important as the population of space debris grows and as missions extend beyond easy communication.6
At a fleet level, advanced concepts explore the use of distributed decision-making, collective learning, and predictive tasking to determine which satellites should be repaired or deorbited first. Similar closed-loop control systems are applied to in-space manufacturing and inspection. However, high-risk interventions still rely on human operators, who must first develop trust in these machines.6
The Shape of the Coming Decade
Progress rests on unglamorous engineering. Radiation-tolerant processors, verified software, carefully labeled datasets, and field trials in quarries and desert analogs decide whether a learned model earns a seat on hardware that no one can repair once it leaves the launch pad.4
The direction holds steady across agencies. Perception, localization, and planning keep migrating from control rooms toward the vehicle, one capability at a time, proven first where supervision is cheap. Lunar surfaces provide that proving ground, and the methods validated there will carry the weight of missions much farther out.1
References and Further Reading
- NASA’s Perseverance Rover Completes First AI-Planned Drive on Mars. (2026). NASA JPL. https://www.jpl.nasa.gov/news/nasas-perseverance-rover-completes-first-ai-planned-drive-on-mars/
- Verma, V. et al. (2023). Autonomous robotics is driving Perseverance rover’s progress on Mars. Science Robotics. DOI:10.1126/scirobotics.adi3099. https://www.science.org/doi/10.1126/scirobotics.adi3099
- Miao, Q., & Wei, G. (2025). A Comprehensive Review of Path-Planning Algorithms for Planetary Rover Exploration. Remote Sensing, 17(11). DOI:10.3390/rs17111924. https://www.mdpi.com/2072-4292/17/11/1924
- Luna, C. et al. (2025). AI-Enabled Capabilities to Facilitate Next-Generation Rover Surface Operations. Research Gate. DOI:10.48550/arXiv.2510.05985. https://www.researchgate.net/publication/396291110_AI-Enabled_Capabilities_to_Facilitate_Next-Generation_Rover_Surface_Operations
- Rachel Hoover. (2023). Part 2: Artificial Intelligence and NASA’s First Robotic Lunar Rover. NASA. https://www.nasa.gov/blogs/missions/2023/12/14/part-2-artificial-intelligence-and-nasas-first-robotic-lunar-rover/
- Levchenko, I. et al. (2026). On-orbit servicing as a future accelerator for small satellites. Npj Space Exploration, 2(1), 14. DOI:10.1038/s44453-025-00024-7. https://www.nature.com/articles/s44453-025-00024-7
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.